Since these models are pretty similar, we should try to make them consistent with what they log to stdout and to Tensorboard. Added Tensorboard support for imagenet in https://github.com/pytorch/xla/pull/985 and we should add the same functionality for MNIST and Cifar10